Delivering end-to-end statistical QoS guarantees for expedited forwarding

https://doi.org/10.1016/S0140-3664(00)00355-8Get rights and content

Abstract

This paper presents an admission control framework for Expedited Forwarding (EF) traffic in a Differentiated Service Network. The aim is to overcome the limitations, in terms of achievable efficiency, which are proper of a deterministic “worst-case” approach based on the zero-loss assumption. An admission control procedure is defined which provides quantifiable end-to-end QoS guarantees in terms of maximum delay and per-flow loss probability. The admission control scheme relies on the analytical derivation of a bound for the per-flow loss probability at a generic network node. The analytical approach is based on the insertion of a discarding device before the EF queue. The purpose of the dropper is to discard packets in order to avoid conflicts at burst scale in the queue, and allows for simple analytical handling of the per-flow loss process. The degradation of the statistical characteristics of the flow along its path are taken into account.

Finally, a comparison between analytical bounds and actual performance results obtained by simulations is presented. The results show that the requested QoS targets are largely met and that the achievable efficiency is much higher than that derived from the worst-case allocation.

Introduction

At present a growing number of Internet applications require some kind of Quality-of-Service (QoS) guarantees, such as delay constraints or low packet loss. In the framework of the actual Internet architecture only the best effort service can be provided, so that additional mechanisms have to be defined in order to bring QoS into the Internet [5]. In this context, the Differentiated Services (DiffServ) model is under study within the IETF [1], [2]. The DiffServ model is aimed at providing QoS on a per-aggregate basis: a limited set of service classes (called PHB in the DiffServ terminology) is supported, each one being associated to a specific QoS level. The internal routers handle packets according to a PHB identifier (DS code), and do not distinguish the individual flows. Then resources are allocated on a per-class basis. In the DiffServ model the functional complexity is mostly located at the network edge, where traffic control functions have to be enforced as admission control, traffic policing and/or shaping, packets marking, etc.

The set of PHB to be implemented is currently under discussion, at present only two PHB have been defined: Expedited Forwarding (EF) [3] and Assured Forwarding (AF) [4]. This paper deals with EF. The EF PHB is intended for supporting real-time applications and in general applications with stringent requirements in terms of delay and jitter and a strictly controlled peak emission rate. It is also to be used to provide the so-called Virtual Leased Line (VLL) service. In other words, it should be considered as the ‘top’ traffic class. Inside the network, the packets marked EF should be served with higher priority over non-EF ones, so that competition versus lower classes is minimal. On the other hand, in order to deliver the target QoS, the network manager has to pro-actively control the input traffic by means of an opportune flow admission control (FAC) enforced by policing/shaping actions at the ingress point.

Defining a suitable FAC scheme for EF is an actual major topic of interest in the DiffServ area: it should be able to ensure the target end-to-end QoS to the admitted flows, but at the same time it should avoid unnecessary under-loading of the network. Any QoS driven admission control relies on the possibility of effectively evaluating the involved QoS performance parameters given the network load state, whose characterization depends on the adopted input traffic parameters. Note that evaluating end-to-end QoS performance parameters in a multistage network is generally a more challenging task than doing the same at a single multiplexer: in fact, in the network each multiplexing stage has an impact on the statistical characteristics of the flows through it, and such an impact has somehow to be taken into account in the evaluation of the multiplexing process at the successive stages.

The traditional approach to the problem of EF allocation follows a ‘deterministic worst-case’ analysis strategy: in a deterministic zero-loss scenario (infinite buffers) the deterministic maximum delay is investigated by the analysis of worst-case arrival patterns. The deterministic maximum delay is assumed as the only QoS parameter and is used for the dimensioning of queues and play-out buffers. In order the deterministic maximum delay to be limited to a reasonably value, the EF traffic load must kept low. This approach, in general, leads to an utilization efficiency much smaller than that theoretically acceptable, as in a zero-loss scenario the deterministic maximum delay is largely faraway from any percentile of practical interest, given a non-trivial level of flow aggregation.

As an alternative, we envisage a scenario where a little loss probability is allowed for EF, provided that such a loss probability can be quantified and controlled. Allowing for a little loss probability means shifting from deterministic toward statistical QoS guarantees, and would permit to push up the achievable mean utilization efficiency by exploiting the statistical characterization of the traffic flows — typically their activity — if they are known and controllable. Obviously the introduced loss probability must be kept below the threshold tolerated by applications. We remark that such an alternative scenario is specially suitable for — but not limited to — Voice over IP applications, which present looser loss requirement and for which a well characterized traffic model (markovian) is available.

In a non-zero loss scenario, this paper is aimed at providing an admission control scheme for EF able to guarantee a quantifiable end-to-end QoS level in terms of maximum delay and packet loss probability. The underlying analytical approach is presented. It takes into account the sources activities and can handle a heterogeneous scenario with different traffic sources. The procedure is based on the evaluation of an analytical bound for the packet loss probability experienced by a flow at a generic node along its path. This bound is determined as a function of the average and peak packet rates of the competitive flows. The degradation of the statistical characteristics of any flow through the network is opportunely taken into account. The typical approaches to statistical admission control described in the literature ([7], [14] provide comprehensive surveys) fail to take into account this degradation and limit their analysis to a single stage multiplexer, so that they are not able to provide any end-to-end QoS guarantee in a multistage network.

The analytical derivation of the a bound for the loss probability at the generic multiplexer relies on the insertion of a packet discarding module, called ‘dropper’, before the FIFO queue. The purpose of the dropper is to discard packets in order to avoid conflicts at the burst scale, and allows for a simple analysis of the loss process at the successive FIFO queue. The loss possibly introduced by the dropper is properly taken into account in the total loss budget.

An extensive performance study has been carried out. The comparison between analytical and simulation results shows that the requested QoS targets are largely met. Section 2 presents the general framework of the proposed admission control mechanism. In Section 3 the underlying analytical approach is discussed and simulative validation is presented. Section 4 comments on the implementation aspects. Finally, in Section 5 the performances results are discussed for a sample scenario.

Section snippets

A flow admission control framework for EF service

The FAC function can be logically assumed to be located in a centralized entity (called Bandwidth Broker, BB [8]). As suggested in Ref. [13], we assume that at every node the EF packets are scheduled with non-preemptive priority over non-EF ones. The impact of non-EF traffic can be considered as an impairment to the EF packets, due to the additional delay possibly caused by the non-EF packets that are completing transmission when an EF packets enters the node. The effect of this additional

An analytical bound for the per-flow loss probability in a single node

The analytical result derived in [9], [10], dealing with multiplexing of independent On/Off packet streams onto an ATM multiplexer, is the starting point for our work. This result is in turn based on a theorem originally derived by Beneš [11]. Therefore, throughout the paper we refer to it as the “Beneš result”. In Section 3.1 results from [9], [10] are summarized, omitting the complete demonstration and highlighting the stringent conditions of applicability. In Section 3.2 we investigate how

Implementation aspects

The right-hand term of bound (25) represents, for the generic multiplexer j, the expression of the function f(average_ratesj, peak_ratesj, parametersj) introduced in Section 2.

As for the dropper parameters D and NF at the generic network multiplexer, two main scenarios can be envisaged

  • (I) D and NF are statically assigned; thus only flows with Ti<d can be admitted, otherwise the bound f(average_ratesj,…) cannot be computed;

  • (II) D and NF are dynamically updated when new flows are admitted.

In

Performance results

The analytical results derived in the previous sections have been validated by simulations. The simulations have been performed using the Network Simulator [13]. The reference network topology, depicted in Fig. 5, aims at representing a multistage network with both an access and a core section, characterized by different link speeds and flows aggregation levels. The goal is to verify that the proposed approach takes into account properly the effects of packet clumping and flows aggregation, and

Conclusions

In this paper a framework for flow admission control in a Differentiated Services Network using the Expedited Forwarding PHB has been proposed. The aim is to overcome the limitations of the traditional deterministic worst case approach. The fundamental assumption is to remove the constraint of deterministic no loss and to consider quantifiable bounds to the loss probability. This approach is especially suitable to deal with variable bit rate flows like voice calls or aggregated voice traffic.

References (14)

  • S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss, An architecture for differentiated service, IETF RFC...
  • K. Nichols, S. Blake, F. Baker, D. Black, Definition of the differentiated services field (DS Field) in the IPv4 and...
  • V. Jacobson, K. Nichols, K. Poduri, An expedited forwarding PHB, IETF RFC 2598, June...
  • J. Heinanen. F. Baker, W. Weiss, J. Wroclawski, Assured forwarding PHB group, IETF RFC 2597, June...
  • X. Xiao et al.

    Internet QoS: a big picture

    IEEE Network

    (1999)
  • M. Listanti, L. Veltri, Worst case analysis of a cascade of nodes, Coritel Technical Report, 14/99,...
  • E.W. Knightly et al.

    Admission control for statistical QoS: theory and practice

    IEEE Network

    (1999)
There are more references available in the full text version of this article.

Cited by (0)

View full text